Devin Prater @devinprater

‼️POPULAR GAMING OPINION‼️ Being an introvert, staying home to play games, read books and do my nerd hobbies is better than going out and partying.

@mikedoise Hey um, so yeah like, at first it was going 15 tokens per second, which I mean that's how fast 12B it-qat is. But then after it got comfy and wasn't being brought up from cold SSD storage, it goes at 30 per second, or 27 at the slowest, at 32K context! 64K maxes my system.

Holy crap y'all, gemma4:26b-a4b-it-qat runs at 15 tokens per second on this Nvidia 4070, 8 GB VRAM, 16 GB RAM, Windows 11, almost as fast as the 12B model! What sorcery is this? @ollama #ai #ollama

The longer I work in corporate, the more I realize... they will make you feel lucky to have the job while underpaying you to do three jobs.

@Favwontmiss Oh wow. I guess I really am neurodivergent then.

If you’re neurodivergent, you’ll understand this instantly: Rest isn’t rest if you’re still being perceived

I’ve got a friend who was non functional after a few hours with a GI bug. I gently pointed out their symptoms matched the ones I deal with every single day. Their response? “You’re used to it. It’s not the same”. That’s how disabled people are treated in a nutshell.

“Authors, please use easy words in your books” is a sentence designed in a lab to upset me. You only learn by reading things that challenge you. This idea that we should only ever be exposed to things we know already is deeply incurious and anti-intellectual.

Alex R. - Fantasy Writer @alexr_writer

2 weeks ago

English is my L2, yet I know what "scrying" means thanks to fantasy. Authors use precise words because worldbuilding demands precision. There’s no shame in not knowing a word, learning it is easy. But complaining because a book asks you to expand your vocabulary? Pure laziness.

65 207 2K 69K 43

So I've never been great at creating video content, so I typically write instead. I'd like to invite everyone to check out my Substack and subscribe over at michaeldoise.com where I write about AI and other technology.

@mikedoise Hmm, oh yeah, I could probably hook Perspective Intelligence up to Ollama on the local network too. That'd be cool.

@mikedoise lol I'd have to get Windows to hunker down and give me that RAM, that Microsoft PC Manager thing works well for that. Maybe 5 tokens per second? I'm still amazed that e4b I can push to 100k context context without it slowing down.

Working on a very long video reading every single codex entry from Mass Effect 3 and it's reminding me how much I love the Mass Effect universe. Need 500 more stories in it 😭

You know, I think AI development is wrong. I mean, it is great, but we are missing the point. AI development should be where anyone can pick up a Mac and a PC and describe what they want and get a fully featured App that uses the latest SDKs and APIs. Describe Shortcut is probably the best step in the right direction for that.

So I stopped waiting and built it in two halves: accessibility agents that review your code as you build (open source), and teardowns, affordable human screen reader testing. The full piece: taylorarndt.substack.com/p/the-state-of… The agents: github.com/Community-Acce… A teardown: taylorarndt.gumroad.com/l/teardown

Excellent video. I am blind and I’m using my iPhone every day to have a life as normal as possible. I’m super happy with all the new features in iOS 27, including the new Siri AI intelligent image descriptions and explore screen. This is a life changer for blind users.

Joanna Stern @JoannaStern

2 weeks ago

Siri is good. Like good-good. After years of broken promises, Apple’s rebuilt the foundation it needs for the future. New YouTube video about my week with Siri AI: youtube.com/watch?v=uQ3lL6…

30 46 516 41K 103

Privacy AI 2.1.1 is live FEATURES - Smart Chat Automatic Model Routing Smart Chat picks the best AI model for every message automatically, so you never have to choose. Set up your model tiers once (a fast model for simple questions, a capable model for complex work, a reasoning model for hard problems, and a local model for anything private), then just type. Smart Chat reads each message for complexity, privacy sensitivity, and content type and routes it, with a badge on each reply showing which model answered and why. Personal or sensitive messages stay on a local on-device model, you can set per-setup spending budgets, and you can override any single choice. - Video Input For Local Vision Models Attach a video to a chat with a local vision model (such as Qwen-VL or Gemma) and it describes what happens across the clip. The app samples evenly spaced frames and feeds them to the on-device model, so it sees motion and multiple moments rather than a single still. This matches cloud models, fully on device and private. IMPROVEMENTS - Faster Sharing To Chat Sharing a file such as a video now opens the chat right away and prepares the file inside it, with a progress overlay, before sending automatically. No more waiting on a loading screen. If the file cannot be processed, the chat still opens and explains what went wrong. Works in regular chats and Smart Chats. - Resume Interrupted Model Downloads Large model downloads now pick up where they left off instead of restarting from zero. If your connection drops, you cancel and restart, or you leave the app and return, the download continues from where it reached, saving time and data on multi-gigabyte files. Applies to all model types. - Resend Keeps Both Replies Resending now keeps the original AI reply and adds the new one as a variant instead of replacing it. An inline picker (such as 2 of 3) appears on the latest reply so you can flip between versions and pick one. Switching a variant also updates what the assistant carries forward. Works in main chat and Smart Chats. - Local Engine Update (llama.cpp b9663) The on-device inference engine was updated to the latest build, adding video understanding for multimodal models, faster image processing when several pictures are attached, new model architectures, and audio and graphics fixes. - MLX Engine Update (Gemma 4 and Vision) The Apple Silicon MLX engine was updated for faster, steadier on-device models. Gemma 4 gains optional speculative decoding for quicker replies, and its image understanding uses less memory so long prompts with pictures no longer fail on iPhone and iPad. Qwen vision uses less memory too, and Gemma 4, Falcon H1, and LFM2 are more accurate. BUG FIXES - MiniMax-M3 Image and Video Input Fixed multimodal input for MiniMax-M3. Images no longer fail with an "invalid image detail" error, and attached videos are now sent in a format the model recognizes, so it describes the video instead of ignoring it. Videos over 50 MB show a note suggesting the convert-to-text option.

@tayarndt yeah, i bought it because i want to know more about the best ways to use a macbook pro, with voiceover.

I wrote a book. My first one. For the first time in a long time there's a Mac under 600 dollars, and Getting Started with the MacBook Neo walks you through setting one up, every step, start to finish, like I'm at the desk with you. Chapter one is free. The price only goes up as I add chapters, so today is the cheapest it will ever be. And if you know someone who'd like it, share it. Book: taylorarndt.gumroad.com/l/macbook-neo Free chapter: open.substack.com/pub/taylorarnd…

Closet Ideas for Easy Accessible ID for Shoes, Boots & Other Footwear When Blind — Part 1 youtu.be/Il7MOvQVu2I